fix(local_calculate_verdicts): retire the border-mismatch join-key veto in favour of a demoted vote

[WilfordGrimley]

Summary

• Retires the border-mismatch join-key veto (local_calculate_verdicts._apply_agreement_checks): a border disagreement between ImageEvidence.layout_class and the matched printing's CanonicalPrintingMetadata.border_color no longer withholds the match (JOIN_KEY_BORDER_MISMATCH_SKIP_REASON). Owner ruling: this catalogue indexes proxy renders of a printing, not the printing itself, so a differently-bordered render still depicts the same printing - "a printing matches exactly, a border mismatch is an accurate distinction."
• The vote is now cast at a new demoted-confidence tier, JOIN_KEY_CONFIDENCE_BORDER_DISAGREEMENT = 0.70, placed (reasoned, not calibrated) between JOIN_KEY_CONFIDENCE_ARTIST_DISAGREEMENT (0.65) and JOIN_KEY_CONFIDENCE_COLLECTOR_ONLY/JOIN_KEY_CONFIDENCE_SYMBOL_TIEBREAK (0.75). When a card also carries an artist-OCR disagreement, _apply_agreement_checks now takes the LOWER of the two tiers via min(...) rather than whichever check runs last silently overwriting the other's demotion.
• The frame half of local_identify_printing_tags.printing_attribute_disagreement is unchanged and still a hard veto (JOIN_KEY_FRAME_MISMATCH_SKIP_REASON) - only the border half changes, per spec.
• JOIN_KEY_BORDER_MISMATCH_SKIP_REASON is retired-not-removed, matching the JOIN_KEY_PROXY_MARKER_VETO_SKIP_REASON precedent this file already established (2026-07-21): declared, still a member of JOIN_KEY_NO_HIT_SKIP_REASONS, so historical rows continue routing to human review via question_feed/the slow-path calculator/the funnel "withheld" count in views.py. No code path writes it as a NEW value going forward.
• docs/reference/skip-reasons.md's border-mismatch row updated in place to reflect the retirement (mirroring the proxy-marker-veto row's own shape), with an explicit note that a stale pre-2026-08-04 row needs no dedicated retraction command - unlike proxy-marker-veto, a plain re-run of run_join_key_calculator under a fresh run_id (2026-07-29's run-scoping) revisits and casts the demoted vote through the ordinary calculator path.
• Module docstring updated (the "Geometry/border agreement + frame agreement" bullet, the "TWO FURTHER CHEAP ADDITIONS" slow-path routing paragraph, the _apply_agreement_checks ordering list) so the border/frame asymmetry is documented at every place it used to describe both as symmetric vetoes.

Scope items 3 and 4 (border tagging / illustration tagging), investigated, no code change needed:

• The observed border is already cast as its own attribute-chip tag by local_layout_class_cast.run_layout_class_cast, wired into Stage E's _run_attribute_chip_casters (2026-07-30) - its own eligibility query is independent of the join-key calculator entirely (it excludes only cards it has itself already voted on), so this was never gated behind the border veto and needs no new caster.
• Investigated illustration tagging for the border-mismatch cohort (12,442/15,626 cards carry artist_ocr_name, per the brief's own measurement). Finding: local_illustration._eligible_illustration_cards_queryset scopes its population to cards the join-key calculator concluded have "no confident hit" (join_key_voted_card_ids = is_no_match=True votes, OR join_key_scanned_card_ids = a JOIN_KEY_NO_HIT_SKIP_REASONS scan-log row). Once the border veto is lifted, a border-mismatched card gets a real, positive join-key vote instead - it is no longer part of that "no confident hit" population, and so it now falls OUT of the illustration calculator's eligibility rather than into it. This is a negative finding, not a gap: these cards are "already tagged" once the veto is lifted, just via the direct join-key printing vote from this PR's own item 1, not via a separate illustration vote. No further change made for this item, per the brief's own instruction to report an accurate negative finding rather than build a fix for a problem that isn't there.

Test plan

• TestAgreementChecks::test_border_mismatch_casts_the_vote_at_a_demoted_confidence (new) - a border disagreement now casts the join-key vote at JOIN_KEY_CONFIDENCE_BORDER_DISAGREEMENT instead of withholding it.
• TestAgreementChecks::test_border_and_artist_disagreement_together_take_the_lower_confidence (new) - both disagreements present -> the lower (artist, 0.65) tier wins, not whichever check ran last.
• TestAgreementChecks::test_frame_mismatch_withholds_the_match (pre-existing, unmodified) - confirms the frame veto is unaffected.
• TestAgreementChecks::test_border_mismatch_writes_a_demoted_vote_via_the_full_runner (rewritten from the old scan-log-row assertion) - the real run_join_key_calculator batch runner casts a CardPrintingTag at the demoted confidence and writes no CardScanLog row for it; a same-run_id re-run does not reconsider the card (ordinary vote idempotence).
• TestFrameVetoRequiresArtistOcr::test_the_border_demotion_is_unaffected_by_the_frame_gate (rewritten from the old veto assertion) - the frame-gate's artist_ocr requirement is scoped to the frame check alone; border still demotes independent of it.
• cardpicker/tests/test_local_calculate_verdicts.py run individually: 187 passed.
• cardpicker/tests/test_local_calculate_verdicts.py + test_skip_reason_roster.py + test_run_scoped_eligibility.py + test_local_identify_printing_tags.py + test_local_illustration.py + test_local_layout_class_cast.py + test_question_feed.py run together (issue Test suite is order-dependent: leaked fetch-failure window trips the envelope across files (8 failures on master) #679 order-dependence check): 680 passed.
• test_skip_reason_roster.py, test_run_scoped_eligibility.py, test_local_identify_printing_tags.py, test_local_illustration.py, test_local_layout_class_cast.py run individually together (no funnel_counts): 447 passed - these files construct border-mismatch CardScanLog rows manually to test historical-row handling and needed no changes, confirming the retired-not-removed treatment is correct.
• pre-commit run on the touched files (ruff, isort, black, mypy, prettier, readme-parity) - green, prettier auto-reformatted the docs/reference/skip-reasons.md table (re-committed).
• test_funnel_counts.py deferred - 24/25 tests in this file fail with a pre-existing, unrelated ValueError: Missing staticfiles manifest entry for 'cardpicker/favicon.ico' in this environment. Verified this is not caused by this change: identical failure reproduces on unmodified master (git stash + re-run). Not investigated further as out of scope for this task.
• CI - not watched by this session per dispatch convention.

Task-end checks (CLAUDE.md)

• Docs convention: edited docs/reference/skip-reasons.md in place (no dated section appended).
• Policy text dates: N/A, no policy page touched.
• Wiki maintenance: N/A - this changes internal calculator behaviour only, no user-visible or admin-visible surface.
• Extractable-primitives ledger: N/A - this touches Stage D's join-key calculator, which already depends on vote-consensus/CanonicalPrinting-consensus machinery; not a new dependency-free primitive.
• Constant-rename-equivalence script: N/A - no constant was renamed, extracted, or moved. JOIN_KEY_BORDER_MISMATCH_SKIP_REASON keeps its exact name and value; only which code paths reference it changed. JOIN_KEY_CONFIDENCE_BORDER_DISAGREEMENT is a brand-new constant, not an extraction of an existing one.

[WilfordGrimley]

Closing unmerged on owner ruling. A border disagreement means the card is not that printing - the same illustration published with a different border is a different printing - so casting a CardPrintingTag for it is a false positive rather than a weak positive. The demoted confidence tier does not mitigate this, because resolve_weighted_consensus weights votes by source alone and never reads the confidence field, so a 0.70 row carries identical weight to a full-confidence join-key hit. The change is also self-defeating. Casting the vote makes the card no longer a no-confident-hit outcome, which drops it out of the illustration calculator's eligible population and discards the one accurate record obtainable here. The existing border veto is correct behaviour and stays as-is. The real gap is downstream in the illustration calculator and is tracked separately.